A two-stage detection method for sea-surface extended targets based on DK-TBD

Abstract

A kind of DK-TBD-based sea surface extended target two-stage detection method, it is related to sea surface target detection technical field.The present application is to solve the problem of low detection accuracy of the existing sea surface target detection method still exists.The present application includes: the radar data sequence set is preprocessed, obtains the set of all real nodes in each frame radar data sequence;Using all real node sets q k' And q k'+1 Candidate trajectory is generated, candidate trajectory set is obtained, and trusted trajectory set is obtained using candidate trajectory set;Using re-tracking method, the trusted trajectory is corrected, to obtain sea surface target detection result.The present application is used for the detection of sea surface target.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and particularly relates to a two-stage detection method for sea surface extended targets based on DK-TBD. BACKGROUND

[0002] The detection of sea targets has always been a hot topic in the research of marine radar detection systems. Unlike optical detection systems, radar detection systems are easily affected by the complex scattering of the sea surface, resulting in low signal-to-noise ratio of the echo signal and often accompanied by a large amount of clutter.

[0003] Traditional radar systems use signal processing algorithms to obtain target information with high reliability, and one of the most classic methods is the constant false alarm rate (CFAR) detection algorithm. The CFAR method statistically analyzes the energy information around each spatial position echo in a specific way, and adaptively sets the threshold to distinguish whether the echo is from a target. Due to the different resolutions of radars, the detected targets also have different forms. In the range-azimuth coordinate, point targets are usually isolated and scattered, while extended targets are clustered into point clusters. If the radar echo data frame in this coordinate form is regarded as a 2-D single-channel digital image, the extended target will appear as an irregular closed polygon with a certain area. On this basis, combined with some simple digital image processing techniques such as edge detection and masking, the spatial position information and geometric shape information of the extended target can be obtained. Scholars have borrowed the design from optical detection systems and simplified these information into a rectangular bounding box, and the target state has become a combination of the center position and the width and height of the bounding box. Then, for these candidate target states obtained in multiple consecutive data frames, the traditional radar system uses methods based on joint probability data association filter (JPDAF) or Kalman filter (KF) to associate or recursively calculate the candidate target states in different data frames to obtain the target motion trajectory, and this series of processing is called detection before tracking (DBT). In the DBT process, the extraction process of candidate targets has almost a decisive influence on the further data association. A high false alarm rate of candidate targets will cause errors in data association, and reducing its false alarm rate may cause real target detection to be missed, thereby leading to low target tracking accuracy.

[0004] To solve the problems in the process of detecting sea surface targets by DBT method, researchers have developed a class of one-time detection and tracking methods called track-before-detect (TBD). TBD is a method of deducing detection results from tracking trajectories. Its core is the integration and accumulation of target effective information and the threshold screening of reliable trajectories. Similar to the DBT method, the TBD method also has two links: candidate target generation within a single data frame and target association between multiple data frames. The difference is that the TBD method accumulates the test statistics (such as square amplitude and envelope likelihood ratio) with high discrimination between target echoes and clutter, and designs appropriate thresholds to distinguish between real trajectories and false alarm trajectories to obtain detection and tracking results for multiple data frames. Due to the accumulation of information between multiple data frames, the sea surface target detection method based on TBD can better handle the problem of weak target echo energy caused by low signal-to-noise ratio, and has a certain robustness to possible missed detections and errors in data association. Researchers have integrated and improved some classic basic theories (such as Hough transform, particle filter, and dynamic programming DP) with TBD to propose a series of effective methods. However, the performance of these basic methods has declined to some extent when migrating from simple scenarios to complex environments. Among them, the DP-TBD method has received special attention due to its strong adaptability to complex environments and the absence of a pre-set target motion model. L. Gao, X. Li, M. Wang, H. Chen, Z. Sun and G. Cui, "Joint Intra-Frame and Inter-Frame Integration Method for High Speed Weak Target Detection," 2022 7th International Conference on Signal and Image Processing (ICSIP), Suzhou, China, 2022, pp. 290-294, doi: 10.1109 / ICSIP55141.2022.9886589. To address the information accumulation error problem caused by distance walking in the TBD process, the peak energy after Radon-Fourier transform (RFT) is used to improve the test statistics of DP-TBD.Haichao Jiang, W. Yi, Guolong Cui, L. Kong and Xiaobo Yang, "Radar detection and tracking of targets in the presence of clutter edge via DP-TBD," 2015 IEEE Radar Conference (RadarCon), Arlington, VA, USA, 2015, pp. 0339-0343, doi: 10.1109 / RADAR.2015.7131021. The impact of the heterogeneous Rayleigh distribution clutter on the use of signal amplitude or log-likelihood ratio as test statistics in DP-TBD is studied, and an improved scheme is proposed by adding a data preprocessing module for clutter suppression before the TBD framework. L. Wang, G. Zhou and T. Kirubarajan, "Track-Before-Detect Method Based on Preprocessing with Pseudo-Spectrum," 2020 IEEE Radar Conference (RadarConf20), Florence, Italy, 2020, pp. 1-6, doi: 10.1109 / RadarConf2043947.2020.9266669. The loss of algorithm performance caused by data preprocessing is considered, and the radar data frame is regarded as an image and a pseudo-spectrum is constructed for each candidate target, and the peak value of the accumulated pseudo-spectrum is used to improve the algorithm performance.E. Grossi, M. Lops, and L. Venturino. A novel dynamic programming algorithm for track-before-detect in radar systems. IEEE Trans. Signal Process., vol. 61, no. 10, pp. 2608-2619, May 2013. Aprile, E. Grossi, M. Lops, and L. Venturino. Track-before-detect for sea clutter rejection: Tests with real data. IEEE Trans. Aerosp. Electron. Syst., vol. 52, no. 3, pp. 1035-1045, Jun. 2016. and W. Yi, L. Kong, and J. Yang. Thresholding process based dynamic programming track before-detect algorithm. IEICE Trans. Commun., vol. E 96. B, no. 1, pp. 291-300, 2013. The low threshold CFAR combined with image segmentation is used to generate candidate targets, and then the DP-TBD algorithm is used to generate track. However, the above-mentioned improved method of DP-TBD mainly studies the adaptability of data preprocessing method to clutter environment and the discrimination degree of test statistics to clutter and target echo. The DP-TBD process mainly includes three parts of candidate target (also called node) generation, adjacent target search association and threshold discrimination to form track. Under the condition of low signal-to-noise ratio, there is usually a difference between the candidate extended target state obtained by preprocessing and the true state, such as the center point position deviation and the scale reduction of the surrounding frame. In the next stage, the target may fail to be associated due to missed detection. These factors cause the accumulated test statistics to be lower than the reasonable threshold, and part of the target detection fails, and the successfully detected target appears a certain error, so the current sea surface target detection method still has the problem of low detection accuracy. SUMMARY

[0005] The purpose of the present application is to solve the problem of low detection accuracy of the existing sea surface target detection method, and a two-stage detection method for sea surface extended target based on DK-TBD is proposed.

[0006] The two-stage detection method for sea surface extended target based on DK-TBD is specifically:

[0007] Step one, pre-process the radar data sequence set X 1:N ={x1,x2,...,x N N} to obtain the set Q 1:N ={q1,q2,...,q k ,...,q N} composed of all real nodes in each frame of radar data sequence.

[0008] Wherein, x N is the Nth frame of radar data sequence, q N is the real node set in the Nth frame of radar data sequence, k∈[1,N], N is the total frame number of radar data sequence, k is the frame number of radar data sequence, q k is the real node set in the kth frame of radar data sequence.

[0009] Step two, generate candidate trajectories by using all real node sets q k' and q k'+1 in adjacent two frames of radar data sequence, obtain the candidate trajectory set T={t1,...,t j' ,...,t s}, and obtain the credible trajectory set

[0010] Wherein, k'=1,...,N-1, q k' is the real node set in the k' frame of radar data sequence, q k'+1 is the real node set in the k'+1 frame of radar data sequence, j'∈[1,s], t j' is the j'th candidate trajectory, j' is the candidate trajectory number, s is the total number of candidate trajectories, is the s'th credible trajectory, s' is the total number of credible trajectories.

[0011] Step three, correct the credible trajectories by using the re-tracking method, so as to obtain the sea surface target detection result.

[0012] The beneficial effects of the present application are:

[0013] The method DK-TBD (DP-KCF-TBD) provided by the application combines target rough detection and re-tracking. The application first generates candidate targets (real nodes) by using constant false alarm detection and a drawing segmentation method for each data frame, generates virtual candidate associated targets (virtual nodes) of isolated targets in the next frame of radar data sequence by using a KCF method, then fuses nodes with spatial overlap, and connects nodes of continuous data frames to generate candidate trajectories. The application screens the connected candidate trajectories based on length and cumulative detection statistics, obtains reliable trajectories, and continuously predicts the reliable trajectories again to correct trajectory errors (i.e., re-tracking), and obtains the final sea surface target detection result. The application reduces the missed detection condition and improves the sea surface target detection accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The flowchart of the application is shown in the figure;

[0015] Figure 2 The main step schematic diagram of the application is shown in the figure;

[0016] Figure 3(a) is a simulation target motion estimation trajectory diagram under the Rayleigh clutter background;

[0017] Figure 3(b) is a simulation target motion estimation trajectory diagram under the K clutter background;

[0018] Figure 4(a) is a PSNR-P curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the Rayleigh distribution clutter background; d

[0019] Figure 4(b) is a PSNR-P curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the Rayleigh distribution clutter background; d

[0020] Figure 4(c) is a PSNR-P curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the Rayleigh distribution clutter background; d

[0021] Figure 4(d) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the Rayleigh distribution clutter background;

[0022] Figure 4(e) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the Rayleigh distribution clutter background; ​​​

[0023] Fig. 4(f) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 in the Rayleigh distributed clutter background;

[0024] Fig. 4(g) is a PSNR-MIoU curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 in the Rayleigh distributed clutter background;

[0025] Fig. 4(h) is a PSNR-MIoU curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 in the Rayleigh distributed clutter background;

[0026] Fig. 4(i) is a PSNR-MIoU curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 in the Rayleigh distributed clutter background;

[0027] Fig. 5(a) is a PSNR-P d curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 in the K distributed clutter background;

[0028] Fig. 5(b) is a PSNR-P d curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 in the K distributed clutter background;

[0029] Fig. 5(c) is a PSNR-P d curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 in the K distributed clutter background;

[0030] Fig. 5(d) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 in the K distributed clutter background;

[0031] Fig. 5(e) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 in the K distributed clutter background;

[0032] Fig. 5(f) is a PSNR-RMSE curve comparison diagram of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 in the K distributed clutter background;

[0033] Figure 5(g) shows a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 under the background of K-distribution clutter.

[0034] Figure 5(h) shows a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 under the background of K-distribution clutter.

[0035] Figure 5(i) shows a comparison of PSNR-MIoU curves of different methods under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 under the background of K-distribution clutter.

[0036] Figure 6(a) shows the impact of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 0 against the background of Rayleigh clutter.

[0037] Figure 6(b) shows the impact of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 1 against the background of Rayleigh clutter.

[0038] Figure 6(c) shows the impact of the cumulative number of frames on the performance of the DK-TBD algorithm under the condition that the RCS fluctuation of the target follows the Swerling model of type 3 in the background of Rayleigh clutter.

[0039] Figure 7(a) shows the PSNR-P DK-TBD method with and without repetitive tracking steps under the condition that the target's RCS fluctuations follow a Swerling model of type 0. d Curve comparison chart;

[0040] Figure 7(b) shows the PSNR-P of the DK-TBD method with and without repetitive tracking steps under the condition that the target's RCS fluctuations follow a Swerling model of type 1. d Curve comparison chart;

[0041] Figure 7(c) shows the PSNR-P DK-TBD method with and without repetitive tracking steps under the condition that the target's RCS fluctuations follow a Swerling model of type 3. d Curve comparison chart;

[0042] Figure 7(d) Comparison of PSNR-MIoU curves of DK-TBD method with and without retracking steps under the condition that the target's RCS fluctuation follows the Swerling model of type 0.

[0043] Figure 7(e) is a comparison diagram of PSNR-MIoU curves of the DK-TBD method with or without the re-tracking step under the condition that the RCS fluctuation of the target follows the Swerling model of type 1;

[0044] Figure 7(f) is a comparison diagram of PSNR-MIoU curves of the DK-TBD method with or without the re-tracking step under the condition that the RCS fluctuation of the target follows the Swerling model of type 3. DETAILED DESCRIPTION

[0045] Specific implementation one: as shown in the figure, the specific process of the two-stage detection method of the sea surface extended target based on DK-TBD in the embodiment is as follows: Figures 1-2

[0046] Step one, pre-processing the N-frame radar data sequence set X 1:N ={x1,x2,...,x N} containing multiple extended target echoes, to obtain a set Q 1:N ={q1,q2,...,q N} composed of all real nodes in each frame of radar data sequence, specifically:

[0047] The pre-processing of the N-frame radar data sequence set X 1:N ={x1,x2,...,x N} containing multiple extended target echoes is that the candidate target generated by the pre-processing is called a real node, and the pre-processing specifically includes: cell average constant false alarm rate detection method (CA-CFAR) and drawing segmentation; the threshold of the constant false alarm rate detection is γ1;

[0048] The kth frame of radar data sequence contains a set of m real nodes:

[0049] q k ={n k,1 ,n k,2 ,...,n k,i ,...,n k,m}

[0050] n k,i ={b k,i ,z k,i}={ [x k,i ,y k,i ,w k,i ,h k,i ],z k,i}

[0051] b k,i =[x k,i ,y k,i ,w k,i ,h k,i ]​

[0052]

[0053] where k∈[1,N],N is the total number of radar data sequences, k is the label of radar data sequence, q k is the set of all real nodes in the kth frame of radar data sequence, i∈[1,m], i is the label of real node, m is the total number of real nodes, n k,i is the i th real node in the kth frame of radar data sequence, b k,i is the node space attribute of real node n k,i , z k,i is the average echo power in the node area range Ω where real node n k,i is located, x k,i is the horizontal coordinate of the center point of n k,i , y k,i is the vertical coordinate of the center point of n k,i , w k,i is the width of n k,i , h k,i is the height of n k,i , g k,i (x,y) is the echo intensity at position (x,y) in the kth frame of radar data sequence, (x,y) is any point in the node area range Ω, a k,i is the peak amplitude of the extended target corresponding to the i th real node in the kth frame of radar data sequence, c k,i is the amplitude of the independent and identically distributed clutter of the extended target corresponding to the i th real node in the kth frame of radar data sequence, h(x,y) is the point spread function of the target echo energy dispersed in the preset area.

[0054] This step is to reduce the missed detection of target echo as much as possible. The CFAR method used to generate real nodes usually sets a low threshold γ1, which leads to more false alarms caused by clutter or noise in the real node set q k . The extended target is the circumscribed rectangle of the candidate target.

[0055] Step two, generate candidate trajectories using all real node sets q k' and q k'+1 in any adjacent frame of radar data sequence, obtain candidate trajectory set T={t1,...,t j' ,…,t s}, and obtain trusted trajectory set T' using candidate trajectory set Specifically:

[0056] Step two one, obtain all real node sets q k' and q k'+1 in any adjacent frame of radar data sequence;

[0057] Where k' = 1, ..., N-1;

[0058] Step 22: Initialize the frame number k' of the radar sequence to 1;

[0059] Steps 2 and 3: Initialize the set q consisting of all real nodes in the (k'+1)th frame of radar data sequence. k'+1 The node label i' = 1;

[0060] Step 24: According to the preset distance threshold d nds In q k'+1 Search node n k',i' If a neighboring node is found, proceed to step 25; otherwise, proceed to step 26.

[0061] Where, n k',i' It is the set of real nodes q k' The i'th real node in the array;

[0062] Step 25: Obtain the set of neighboring nodes d k'+1,i' , set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' Connect the trajectories, add the resulting trajectories to the candidate trajectory set, and then proceed to step 27, which is as follows:

[0063] Step 2.51: Obtain n k',i' The set of neighboring nodes d k'+1,i' Then obtain the set of neighboring nodes d. k'+1,i' Detection statistics for each neighboring node:

[0064] λ(n k'+1，j )=z k'+1，j

[0065] Where j is d k'+1,i' The labels of the nearest neighbors, n k'+1，j It is d k'+1,i' The j-th neighboring node, z k'+1，j It is the neighboring node n k'+1，j Average power of the region;

[0066] Step 252: Set d k'+1,i' The neighboring node with the largest detection statistic and node n k',i' The concatenation process involves adding the concatenated trajectory as a candidate estimate to the candidate trajectory set. Specifically:

[0067]

[0068] In this context, → represents a connection.

[0069] Step 26: Generate virtual nodes using the kernel correlation filter method, and then connect the virtual nodes with n. k',i' Connect, and add the virtual node to q. k'+1 Update q k'+1 Then update the Chinese Q k'+1 Overlapping virtual nodes are merged and updated to obtain candidate trajectories, which are then added to the candidate trajectory set. Step 27 is then executed, specifically:

[0070] Step 261, with n k',i' Generate a size w to be tracked centered on. δ ×h δ The initial target template δ k',i' (n k',i' Simultaneously set Gaussian form markers. Using δ k',i' (n k',i' ) and L k',i' Obtain the weight parameter matrix learned by KCF

[0071]

[0072] in, The weight parameter matrix learned by KCF, w δ It is the width of the initial target template, h δ It is the height of the initial target template. It represents the initial template δ k',i' (n k',i' The autocorrelation (kernel) matrix of L k',i' It is the initial response of the target. It is the symbol for a two-dimensional Fourier transform. It is a two-dimensional inverse Fourier transform. It is a positive hyperparameter used to adjust the effect of regularization. It is the kernel function, and τ = 0.5 is the proportionality coefficient of the linear combination. yes The normalization result, yes The normalized result, ⊙, represents element-wise multiplication, σ g It is the standard deviation of the isotropic Gaussian kernel. It is a Gaussian kernel function. It is a kernel function in the form of a dot product, where μ and η are hyperparameters;

[0073] Step 262: Obtain the target template δ in the (k'+1)th frame of the radar sequence. k',i' Monitoring areas of the same size Using the weight matrix get The corresponding response matrix Specifically:

[0074]

[0075] in, It is the initial template δ k',i' With monitoring area The cross-correlation (kernel) matrix,

[0076] Step 263, Utilize The position where the maximum value appears [x] k'+1,j* ,y k'+1,j* Generate virtual nodes:

[0077]

[0078] in, These are the coordinates of the center point of the j*th virtual node generated in the k'+1th frame. It is the width of the j*th virtual node generated in the k'+1th frame. It is the height of the j*th virtual node generated in the k'+1th frame. It is a real node n k',i' Predicted region in frame k+1 average power, It is a node The node status;

[0079] Step 264: Generate virtual nodes With real node n k',i' Connect, add virtual nodes The set of real nodes up to frame k'+1 q k'+1 In the middle, q k'+1 Updated to the following format:

[0080] q k'+1 ={n k'+1,1 ,…,n k'+1,m' ,v k'+1,1 ,…,v k'+1,n}

[0081] Where n is the number of virtual nodes generated, and m' is q k'+1 Total number of nodes;

[0082] Step 265: Update q k'+1The virtual nodes that exist in space are fused to obtain fused virtual nodes, and the virtual nodes that need to be fused are replaced by the fused virtual nodes, so as to obtain a candidate trajectory, and the candidate trajectory is added to a candidate trajectory set;

[0083] The updated q k'+1 is fused:

[0084]

[0085] Wherein, Ψ(··) is a node fusion function, z k'+1,f is the average power of the b k'+1,f region, b k'+1,f is the node state of the fused virtual node, v k'+1,f is the fused virtual node, is the horizontal coordinate of the center of the two virtual nodes that need to be fused, is the vertical coordinate of the center of the two virtual nodes that need to be fused, is the width of the two virtual nodes that need to be fused, is the height of the two virtual nodes that need to be fused.

[0086] Step two seven, judge whether i' is equal to m', if i'=m', execute step two eight; if i'<m', let i'=i'+1 and return to step two four;

[0087] Wherein, m' is the total number of nodes in q k'+1 ;

[0088] Step two eight, judge whether k' is greater than the minimum trajectory length threshold l fn of the target motion, if k'<l fn , let k'=k'+1, return to step two three; if l fn <k'≤N-1, execute step two nine;

[0089] Step two nine, screen the candidate trajectory set to obtain a screened candidate trajectory set; if k'<N-1, let k'=k'+1, return to step two three; if k'=N-1, the screened candidate trajectory set is taken as a set of credible trajectories

[0090] Screening the candidate trajectory set is specifically:

[0091] Obtain the cumulative test statistic λ j' (t t ) of the candidate trajectory t j' , and delete the candidate trajectory corresponding to the cumulative test statistic lower than the preset threshold γ2, to obtain a set of credible trajectories

[0092] candidate trajectory t j' cumulative test statistics λ t (t j' ) is obtained by the following way:

[0093]

[0094] wherein, is the s'th trusted trajectory, s' is the total number of trusted trajectories, is the l'th trusted trajectory, is the j'th node in the j'th candidate trajectory, is the test statistics of

[0095] Step three, re-tracking the trusted trajectories in the trusted trajectory set, correcting the trusted trajectories, and obtaining the sea target detection result, specifically:

[0096] Step three one, taking each virtual node in the trusted trajectory as a real node, connecting all the real nodes to obtain the trusted trajectory, and expressing the trusted trajectory as follows:

[0097]

[0098] wherein, n b is the b'th real node in the trusted trajectory , b is the label of the real node in the trusted trajectory , n b+1 is the b+1'th real node in the trusted trajectory , and L is the number of real nodes in the trusted trajectory ;

[0099] Step three two, using the KCF method to generate an initial template δ' and a corresponding virtual node v'1 using the first real node in the trusted trajectory , then using the initial template δ' and the corresponding virtual node v'1 to obtain a weight matrix, and obtaining a corresponding predicted virtual node for other real nodes, thereby composing a virtual trajectory

[0100]

[0101] An initial template δ' and a corresponding virtual node v'1 are generated using the first real node in the trusted trajectory , then an initial template δ' and a corresponding virtual node v'1 are used to obtain a weight matrix, and the steps of obtaining a corresponding predicted virtual node for other real nodes are the same as steps two six one to two six three.

[0102] ​​Step three, when the trusted trajectory and the virtual trajectory When there is an overlapping part, the center point of the minimum circumscribed rectangle region of the overlapping node region is taken as the center point of the real node in the corrected trusted trajectory, so as to obtain the corrected trusted trajectory, and each node in the corrected trusted trajectory is the target detection result;

[0103]

[0104] Wherein, Ψ(n b ,v b ) is the bth real node in the corrected trusted trajectory.

[0105] The invention proposed in the application is named as DK-TBD algorithm.

[0106] The pseudo code flow of the application is shown in Table 1.

[0107] Table 1

[0108]

[0109] Embodiment:

[0110] In order to verify the beneficial effects of the application, simulation experiments are carried out in this embodiment.

[0111] In this embodiment, the DK-TBD proposed in the application is compared with some TBD methods in simulated data under Rayleigh distribution and K distribution clutter background, including DP-MSA-TBD, DP-WSA-TBD and MKCF-TBD. Among them, DP-MSA-TBD represents the classical DP-TBD method using average square amplitude as the test statistic, DP-WSA-TBD represents the improved DP-TBD method using weighted square amplitude as the test statistic, and MKCF-TBD is a TBD method completely based on KCF. All algorithms in this embodiment are implemented using Python and tested on a computer with i7-12700KF CPU (3.60-GHz frequency) and 32-GB RAM.

[0112] In this embodiment, the following indicators are used to evaluate the tracking results of multiple algorithms before detection.

[0113] (1) Discovery probability P d : When the intersection ratio (i.e. the percentage of the overlapping part area in the sum of the total areas of the two) of the detected extended target and the corresponding real value exceeds a certain threshold, the target is considered to be discovered, and P d represents the ratio of the detected part in the multi-target real trajectory to the sum of the total lengths of all target trajectories, as follows:

[0114]

[0115] where m represents the total number of extended targets in all data frames, and k represents the total number of discovered targets.

[0116] (2) Root Mean Square Error (RMSE): the average distance between the center points of the multi-frame extended target estimation value and the corresponding true value, as follows:

[0117]

[0118] (3) Mean Intersection over Union (MIoU): the average intersection over union of the multi-frame extended target estimation value and the corresponding true value, as follows:

[0119]

[0120] where E and G represent the bounding boxes of the extended target estimation value and the true value, respectively, and S represents the calculated area.

[0121] Then the motion state of the targets in the Rayleigh or K-distributed clutter background is simulated. Table 2 lists the motion parameters of each target.

[0122] Table 2

[0123]

[0124] Four extended targets with different motion models are set: targets A and B move according to the constant velocity motion model (CV), target C moves according to the acceleration motion model CA, and target D moves according to the continuous turning motion model (CT). During the motion process, target A and target D first approach each other and then move away after meeting. The RCS fluctuation of the targets follows the Swerling model of types 0, 1 and 3. The number of data frames in the simulation is set to 50 frames, the resolution of the data frame image is 300x300, and it is assumed that one pixel represents 1 m. The time interval between two frames is 1 s, and the units of initial velocity, acceleration and turning rate are m / s, m / s 2 and ° / s, respectively. The simulation target motion estimation in the Rayleigh clutter background is shown in Fig. 3(a), and the simulation target motion estimation in the K clutter background is shown in Fig. 3(b).

[0125] For the CV motion model, it is assumed that the motion state of the target can be represented by a 4-dimensional vector, which has:

[0126]

[0127] where x k and y k represent the positions of the target in the horizontal direction and the vertical direction, respectively, and denotes the target's horizontal and vertical velocity.

[0128] The state transition matrix of the CV motion model can be expressed as follows:

[0129]

[0130] where I2 denotes a 2x2 identity matrix, Δt denotes the time interval between two frames, and the symbol denotes the Kronecker product.

[0131] The CV motion model can be described as follows:

[0132]

[0133] where is a state vector of zero-mean Gaussian noise.

[0134] For the CA motion model, the motion state of the target can be represented by a 6-dimensional vector, which has

[0135]

[0136] where and denote the target's horizontal and vertical acceleration.

[0137] The state transition matrix of the CA motion model can be expressed as follows:

[0138]

[0139] The CA motion model can be described as follows:

[0140]

[0141] where is a state vector of zero-mean Gaussian noise.

[0142] For the CT motion model, the motion state vector of the target is only related to its position, which has

[0143]

[0144] The state transition matrix of the CT motion model can be expressed as follows:

[0145]

[0146] where R denotes the turning radius and ω denotes the angular velocity of the target's motion.

[0147] The CT motion model can be described as follows:

[0148]

[0149] where the state vector is composed of two zero-mean Gaussian noises and one zero noise.

[0150] According to the parameters given in Table 2, the four targets in the simulation setup move according to the CV model, CA model and CT model described above, respectively. For each simulation frame, the center position and bounding box properties of the four targets are recorded, which are set as the true values to evaluate the performance of the algorithm.

[0151] (2) Extended target RCS fluctuation model setup:

[0152] In the simulation, the Swerling 0, 1 and 3 models are used to simulate the fluctuation of the target, and a two-dimensional elliptical Gaussian function is used to model the diffusion of target energy. The target modeling has:

[0153] g k (x,y|x k )=a k ·h k (x,y|x k )+c k

[0154] where g k (x,y|x k ) represents the value of the pixel point in the extended target x k region, which is also the energy intensity of the point, a k represents the peak amplitude, which changes according to the preset Swerling model, c k represents an independent background clutter, which changes according to the preset clutter distribution.

[0155] The form of h k (x,y|x k ) is as follows:

[0156]

[0157] where x k and y k represent the coordinates of the center position of the extended target x k , the variables x and y represent the relative coordinates of the current pixel point and the center point of x k , σ x and σ y represent the standard deviations of the two axes of the elliptical Gaussian, respectively. θ represents the clockwise rotation angle of the target, which can be obtained through the speed of the target, and there is

[0158] For the peak amplitude a kwhich remains constant in the Swerling-0 fluctuation model, varies according to Rayleigh distribution in the Swerling-1 fluctuation model, and varies according to 4- degree χ 2 distribution in the Swerling-3 fluctuation model.

[0159] The probability density function (PDF) of Rayleigh distribution in the Swerling-1 model is:

[0160]

[0161] The PDF of 4-degree χ 2 distribution in the Swerling-3 model is:

[0162]

[0163] where σ t represents the average power of the target, and σ

[0164] Assuming that the average power of the clutter is constant as σ c , the average peak signal-to-noise ratio (PSNR) between the extended target and the background clutter can be described as:

[0165]

[0166] The average power σ t of the target can be represented by the average power σ c of the clutter and their PSNR, and has:

[0167] σ t = σ c 10 PSNR / 10

[0168] Therefore, for different Swerling types, the peak amplitude of the extended target in the kth frame can be represented as follows: for the Swerling-0 type, the a k of the extended target in all data frames is equal to

[0169] Table 3 lists the main hyperparameter settings of multiple TBD methods. In order to ensure the fairness of the algorithm comparison as much as possible, by controlling the threshold γ2 of the cumulative test statistic and other hyperparameters, the false alarm rate P fa of all methods to detect the track is maintained at about 0.01-0.05.

[0170] Table 3

[0171]

[0172] The embodiments test the detection performance of the present application and various comparative methods under different clutter backgrounds, and evaluate and compare them, wherein the average of 100 Monte Carlo experiment evaluation results is used for each method. The simulation experiment obtains multiple sets of evaluation curves, including PSNR-P d curves, PSNR-RMSE curves and PSNR-MIoU curves, by changing the clutter distribution, PSNR and Swerling type and other conditions.

[0173] Figures 4(a)-4(i) The comparison of the detection and tracking performance of various algorithms under Rayleigh distribution clutter background is shown. In the PSNR-P d graphs (Figs. 4(a), 4(b), 4(c), the black dotted line represents the P d In the conditions of Swerling-0, 1, 3, DK-TBD can obtain higher P d when the PSNR is low (PSNR≤5). d The P d curve of DK-TBD is closer to the black dotted line and may be higher, which reflects the improvement of the virtual node design in DK-TBD on the missed detection in CFAR preprocessing. The P d of the benchmark method is lower than that of DK-TBD, which is because more target losses occur in the tracking process of the benchmark method, which also proves the effect of the virtual node. In the PSNR-RMSE graph (Figs. 4(d), 4(e), 4(f)), the multi-target tracking trajectory of DK-TBD shows a lower RMSE curve. Thanks to the virtual node and re-tracking designs, DK-TBD can have higher accuracy when the PSNR is low. In the PSNR-MIoU graph (Figs. 4(g), 4(h), 4(i)), although the MIoU of DK-TBD is higher than that of the benchmark DP-MSA-TBD method (which completely relies on the target size estimation of CFAR preprocessing), it is not much different from MKCF-TBD at low signal-to-noise ratio, because the multi-candidate frame fusion design in MKCF-TBD greatly improves the size of the detected extended target, but at high signal-to-noise ratio, the overlapping degree of the multiple candidate frames generated by MKCF-TBD becomes high, and the adaptive adjustment of the target size becomes poor, so it becomes low.

[0174] Figures 5(a)-5(i) The comparison of the detection and tracking performance of various algorithms under K distribution clutter background is shown. K distribution clutter is more complex, and more false targets are detected in CFAR preprocessing. The DK-TBD method can still achieve good results in P d , RMSE and MIoU indicators.

[0175] Figures 6(a)-6(c) The cumulative frame number l is analyzed fn The impact on algorithm performance, the discovery probability P of DK-TBD d Without the increase of l fn , even decreased, because greater l fn Make more false alarm nodes mixed into the trusted trajectory, affect the effective connection of the real target node. Especially in the case of low SNR, the false alarm node and the real node gap is smaller, more easily because of the wrong association caused by target missed detection, P d Down. In order to keep a high P d , DK-TBD must be set to a lower threshold γ2, but this will cause the P fa of the detected trajectory to rise. In order to balance P d and P fa , choose to set the cumulative frame number l fn = 3.

[0176] Figures 7(a)-7(f) The effectiveness of the DK-TBD re-tracking part is analyzed, and it can be seen from the figure that under all Swerling types, the DK-TBD method with re-tracking part, P d and MIoU are higher than DK-TBD without using re-tracking, and the DK-TBD method with re-tracking part has lower RMSE, which shows that the re-tracking part can effectively improve the performance of the algorithm.

[0177] In the simulation experiment, compared with other TBD methods, the detection trajectory of DK-TBD has better accuracy. In the case of low SNR, the method proposed in the application can better discover and track the target.

Claims

1. A two-stage detection method for DK-TBD based maritime extended target, wherein the DK-TBD is DP-KCF-TBD, characterized in that The method specifically comprises the following steps: Step one, pre-process the radar data sequence set X 1:N ={x1,x2,...,x N} to obtain the set Q 1:N ={q1,q2,...,q k ,...,q N} composed of all real nodes in each frame of radar data sequence. wherein x N is the Nth frame of radar data sequence, q N is the real node set in the Nth frame of radar data sequence, k ∈ [1, N], N is the total frame number of the radar data sequence, k is the frame label of the radar data sequence, q k is the real node set in the kth frame of radar data sequence; Step two, generate candidate trajectories using all real nodes set q in the two adjacent radar data sequences k' and q k'+1 Generate candidate trajectories, get candidate trajectory set T = {t1, …, t j' , …, t s}, and get trusted trajectory set T using candidate trajectory set where k' = 1,..., N - 1, q k' is the set of real nodes in the k' th frame of radar data, q k'+1 is the set of real nodes in the k' + 1 th frame of radar data, j' ∈ [1, s], t j' is the j' th candidate trajectory, j' is the candidate trajectory label, s is the total number of candidate trajectories, is the s' th trusted trajectory, s' is the total number of trusted trajectories; The all-real node set q in the two adjacent radar data sequences is utilized k' And q k'+1 A candidate trajectory is generated, a candidate trajectory set T = {t1, …, t j' ,…,t s} is obtained, and a trusted trajectory set T' is obtained by utilizing the candidate trajectory set Specifically: Step two, obtaining all real node set q in the adjacent frame radar data sequence k' and q k'+1 ; Wherein, k'=1,...,N-1; Step two, initializing the frame number k' of the radar sequence to 1; Step two three, initialize the set q consisting of all real nodes in the k'+1 frame radar data sequence k'+1 Node index i'=1; Step two four, search for the adjacent nodes of the node n nds In q k'+1 Search for the adjacent nodes of the node n k',i' If the adjacent nodes are found, step two five is executed; if not, step two six is executed. wherein n k',i' is the i'th real node in the set of real nodes q k' ; Step two five, obtaining a neighbor node set d k'+1,i' The neighbor node with the maximum detection statistic in the set d k'+1,i' is connected with the node n k',i' The trajectory obtained after the connection is added to the candidate trajectory set as a candidate trajectory, and then step two seven is executed. Step 26: Generate virtual nodes using the kernel correlation filter method, and then connect the virtual nodes with n. k',i' Connect, and add the virtual node to q. k'+1 Update q k'+1 Then update q k'+1 The overlapping virtual nodes are merged and the virtual nodes are updated to obtain candidate trajectories. The candidate trajectories are then added to the candidate trajectory set, and then step 27 is executed. Step two seven, judging whether i' is equal to m', if i'=m', executing step two eight, if i'<m', setting i'=i'+1 and returning to step two four; where m' is q k'+1 total number of intermediate nodes; Step two eight, judge whether k' is greater than the minimum track length threshold value l of the target motion fn , if k' < l fn , then let k' = k' + 1, return to step two three; if l fn < k' ≤ N - 1, then execute step two nine; Step two nine, screening the candidate trajectory set to obtain a screened candidate trajectory set; if k' < N-1, then let k' = k' + 1, and return to step two three; if k' = N-1, then take the screened candidate trajectory set as the set of trusted trajectories wherein, is the s'th trusted trajectory, s' is the total number of trusted trajectories, is the l'th trusted trajectory; Step three, correcting the credible track by using a re-tracking method, so as to obtain the sea surface target detection result.

2. The two-stage detection method for extended targets on the sea surface based on DK-TBD according to claim 1, characterized in that: The step one in the radar data sequence set X 1:N = {x1, x2, …, x N} is preprocessed: first, the radar data sequence set is processed by using a cell average constant false alarm rate detection method, and then the output of the cell average constant false alarm rate detection method is plotted and segmented to obtain a set composed of all real nodes in each frame of radar data sequence.

3. The two-stage detection method for extended targets on the sea surface based on DK-TBD according to claim 2, characterized in that: The real node set in the kth frame radar data sequence in the step one is specifically as follows: q k = {n k,1 , n k,2 ,..., n k,i ,..., n k,m} n k,i = {b k,i , z k,i} = {[x k,i , y k,i , w k,i , h k,i ], z k,i} b k,i = [x k,i , y k,i , w k,i , h k,i ] wherein i ∈ [1, m], i is a real node label, m is the total number of real nodes, n k,i is the i-th real node in the k-th frame of radar data sequence, b k,i is the node spatial attribute of the real node n k,i , z k,i is the average echo power within the node area range Ω of the real node n k,i , x k,i is the horizontal coordinate of the center point of n k,i , y k,i is the vertical coordinate of the center point of n k,i , w k,i is the width of n k,i , h k,i is the height of n k,i , g k,i (x, y) is the echo intensity at position (x, y) in the k-th frame of radar data sequence, (x, y) is any point within the node area range Ω, a k,i is the peak amplitude of the extended target corresponding to the i-th real node in the k-th frame of radar data sequence, c k,i is the amplitude of the independent and identically distributed clutter of the extended target corresponding to the i-th real node in the k-th frame of radar data sequence, h(x, y) is a point spread function for the dispersion of target echo energy within a preset area.

4. The two-stage detection method for extended targets on the sea surface based on DK-TBD according to claim 3, characterized in that: The step two five in the acquisition adjacent node set d k'+1,i' , the set d k'+1,i' The maximum detection statistics adjacent node and node n k',i' Connect, the trajectory obtained after connection is added to the candidate trajectory set as a candidate trajectory, specifically: Step two five one, obtain n k',i' the neighboring node set d k'+1,i' of the neighboring node set d k'+1,i' , then obtain the detection statistics of each neighboring node λ(n k'+1，j ) = z k'+1，j where j is d k'+1,i' is the label of a neighboring node, n k'+1，j is d k'+1,i' is the jth neighboring node, z k'+1，j is a neighboring node n k'+1，j is the average power of a region; Step two five two, collect d k'+1,i' The node with the largest detection statistic is detected as the neighboring node of node n k',i' Connect, and add the connected track into the candidate track set as a candidate estimate.

5. The two-stage detection method for extended targets over the sea based on DK-TBD according to claim 4, characterized in that: In step two-six, the kernel correlation filtering method is used to generate virtual nodes, and the virtual nodes are then compared with n. k',i' Connect, and add the virtual node to q. k'+1 Update q k'+1 Then update q k'+1 The overlapping virtual nodes are merged and updated to obtain candidate trajectories, which are then added to the candidate trajectory set. Specifically: Step two 261, generate an initial target template δ k',i' (n δ ) of size w δ ×h k',i' to be tracked with n k',i' as the center, and set the label in the form of Gaussian Get the weight parameter matrix learned by KCF using δ k',i' (n k',i' ) and L k',i' ​ wherein, is a weight parameter matrix learned by KCF, w δ is a width of an initial target template, h δ is a height of the initial target template, is an autocorrelation matrix representing the initial template δ k',i' (n k',i' ), L k',i' is an initial response of a target, is a two-dimensional Fourier transform symbol, is a two-dimensional inverse Fourier transform, is a positive hyperparameter, is a kernel function, τ is a proportional coefficient of linear combination, is a normalized result of , is a normalized result of , is an element-wise multiplication, σ g is a standard deviation of an isotropic Gaussian kernel, is a kernel function in a Gaussian form, is a kernel function in a dot product form, μ and η are hyperparameters; Step two 262, obtaining the initial target template δ in the k'th frame radar sequence k',i' Monitoring areas of the same size Using a weight matrix Obtaining A corresponding response matrix Specifically: wherein, is the initial template δ k',i' and the monitoring region is the cross-correlation matrix; Step two 263, utilizing Position of maximum value occurrence Generating virtual nodes Step two hundred sixty-four, the generated virtual node is connected to the real node n k',i' , the virtual node is added to the set of real nodes q k'+1 in the k'+1 frame and q k'+1 is updated: q k'+1 = {n k'+1,1 ,...,n k'+1,m' ,v k'+1,1 ,…,v k'+1,n} where n is the number of generated virtual nodes, m' is q k'+1 the total number of intermediate nodes; Step two 265, update q k'+1 If there is spatially overlapping virtual node fusion in the updated q, the fused virtual node is obtained, and the virtual node to be fused is replaced by the fused virtual node, so as to obtain a candidate trajectory, and the candidate trajectory is added to the candidate trajectory set.

6. The two-stage detection method for extended targets over the sea based on DK-TBD according to claim 5, characterized in that: The step two 263 in using The position of maximum value occurrence Generating a virtual node Specifically: in, These are the coordinates of the center point of the j*th virtual node generated in the k'+1th frame. It is the width of the j*th virtual node generated in the k'+1th frame. It is the height of the j*th virtual node generated in the k'+1th frame. It is a real node n k',i' Predicted region in frame k+1 average power, It is a node The node status.

7. The two-stage detection method of extended targets on the sea surface based on DK-TBD according to claim 6, characterized in that: The step two 265 in the updated q k'+1 There is a spatially overlapping virtual node fusion, specifically: where Ψ (· ·) is a node fusion function, z k'+1,f is the average power of the b k'+1,f region, b k'+1,f is the node state of the fused virtual node, v k'+1,f is the fused virtual node, is the x-coordinate of the center of the two virtual nodes to be fused, is the y-coordinate of the center of the two virtual nodes to be fused, is the width of the two virtual nodes to be fused, is the height of the two virtual nodes to be fused.

8. The two-stage detection method of extended targets on the sea surface based on DK-TBD according to claim 7, characterized in that: The step two nine comprises the following steps: Obtain candidate trajectory t j' The cumulative test statistic λ t (t j' The candidate trajectories corresponding to the cumulative test statistics below the preset threshold γ2 are deleted to obtain the filtered candidate trajectory set. candidate trajectories t j' cumulative test statistic λ t (t j' ) is obtained by wherein, is the j'th node in the j'th candidate trajectory, is a test statistic for .​ 9. The two-stage detection method of extended targets on the sea surface based on DK-TBD according to claim 8, characterized in that: The step three comprises the following steps: The step three comprises the following steps: wherein n b is the bth real node in the trusted trajectory b is the label of the bth real node in the trusted trajectory n is the number of real nodes in the trusted trajectory b+1 is the b+1th real node in the trusted trajectory L is the number of real nodes in the trusted trajectory ​ Step three two, generate initial template δ' and corresponding virtual node v'1 by the first real node in the trusted trajectory , then obtain weight matrix by using initial template δ' and corresponding virtual node v'1, obtain corresponding predicted virtual node for other real nodes, so as to constitute virtual trajectory Virtual trajectory is: Step three, when there is an overlapping part in the trusted trajectory and the virtual trajectory the center point of the minimum circumscribed rectangle region of the overlapping node region is taken as the center point of the real node in the corrected trusted trajectory, so as to obtain the corrected trusted trajectory, and each node in the corrected trusted trajectory is the target detection result; Step three one, taking each virtual node in the credible track as a real node, connecting all the real nodes to obtain the credible track, as follows: The corrected credible track is as follows: wherein Ψ(n b ,v b ) is the bth real node in the corrected trusted trajectory.

Images